Kickstart Conversational Analytics agents with the Looker ChromeUX Block

Key Concepts

• Conversational Analytics Agents: AI-powered interfaces within Looker that allow users to query data using natural language.
• Chrome UX (CrUX) Report: A public dataset containing real-world user experience metrics for websites, dating back to 2017.
• Looker Blocks: Pre-built LookML assets that provide pre-modeled data, metrics, and dimensions, allowing for immediate analysis without writing custom SQL.
• Guardrails: Specific instructions provided to an AI agent to restrict its scope, enforce data quality, and ensure it adheres to business logic.
• P75 LCP (Largest Contentful Paint): A core web vital metric representing the 75th percentile of real user experiences, used to measure site loading performance.

Prerequisites and Setup

Before building an agent, ensure the following environment requirements are met:

1. BigQuery Connection: A default connection to the Chrome UX report project must be established in Looker.
2. Marketplace Access: Permissions to access the Looker Marketplace to install pre-built blocks.
3. User Roles: The creator must have the conversational analytics agent manager role. End-users require the conversational analytics user role.

Installation Process:

• Navigate to the Looker Marketplace, search for "Chrome UX," and install the block.
• Configure the block to use the established BigQuery connection. This provides a pre-modeled dataset, eliminating the need for manual LookML or SQL coding.

Building the Conversational Agent

The process of creating an agent is streamlined into a few logical steps:

1. Initialization: Navigate to the Conversations menu in Looker and select New Agent.
2. Configuration: Provide a descriptive name and a clear purpose (e.g., "Website Performance Reviewer").
3. Data Mapping: Select the appropriate "Explore." The video recommends the Device Summary explore, noting that the "Country Normalized" explore is often too complex for natural language processing.
4. Defining Instructions (Guardrails): This is the most critical step for ensuring accuracy. Examples of instructions provided:
   • Scope Control: Instruct the agent to reject off-topic queries (e.g., weather or poetry).
   • Metric Standardization: Force the use of the P75 LCP field for all speed-related queries and require the agent to explain what this metric represents.
   • Temporal Constraints: Limit data retrieval to January 2024 onwards.
   • Data Formatting: Instruct the agent to handle URL prefixes automatically.

Testing and Validation

Testing is essential to ensure the agent is "grounded" in the provided instructions rather than relying solely on its general training data.

• Negative Testing: Attempt to ask the agent off-topic questions to verify that the rejection guardrails are functioning.
• Constraint Testing: Query data outside the allowed date range (e.g., 2021–2022) to confirm the agent acknowledges the data exists but adheres to the imposed temporal restrictions.
• Comparative Analysis: Use the agent to compare two specific entities (e.g., Google Docs vs. Google Scholar) to verify that the agent can successfully synthesize and present comparative performance metrics.

Best Practices for Robust Agents

• Clarity in Instructions: Use specific, unambiguous language when defining guardrails.
• Labeling and Synonyms: Apply clear labels and define synonyms to help the agent interpret user intent accurately.
• Data Quality: Understand the underlying dataset thoroughly to write better instructions.
• Iterative Experimentation: Continuously test the agent with various prompts to identify edge cases and refine the instructions accordingly.

Synthesis

Building a conversational analytics agent in Looker is a rapid process—achievable in approximately five minutes—when leveraging pre-built Looker Blocks like the Chrome UX report. The effectiveness of these agents relies heavily on the quality of the "instructions" provided during setup. By implementing strict guardrails, defining key metrics (like P75 LCP), and setting clear temporal and topical boundaries, users can transform complex datasets into accessible, high-impact insights for non-technical stakeholders.